[llama4] Add weight-only int4 linear quantization for backbone export#20644
Open
SS-JIA wants to merge 2 commits into
Open
[llama4] Add weight-only int4 linear quantization for backbone export#20644SS-JIA wants to merge 2 commits into
SS-JIA wants to merge 2 commits into
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20644
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 7280475 with merge base 9f31912 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
SS-JIA
added a commit
that referenced
this pull request
Jun 30, 2026
Pull Request resolved: #20644 Adds a `--weight-only-linear` export option that quantizes the TISO backbone's linears to weight-only int4 (dequantized int4 weights, fp activations) instead of the default 8da4w (dynamic int8 activations + int4 weights). Adds a `WeightOnlyInt4Linear` module and `_replace_linear_with_linear_int4_weight_only_for_pre_quantization` in `examples/models/llama/source_transformation/pre_quantization.py`. It uses the same buffer layout as `Int8DynActInt4WeightLinear` (int8 `weight` holding int4 values, per-group `scales` / `zeros`), so an existing QAT checkpoint loads unchanged; its forward dequantizes the weight and runs a plain `F.linear` with no per-token activation quantization. `transform_linear_for_pre_quantization` gains a `weight_only` flag, threaded through the llama4 backbone build (`backbone/model.py`) and exposed as `--weight-only-linear` on `export_llm_backbone.py`. Motivation: the 8da4w dynamic-activation-quant path triggers a Mali-GPU-specific numerical bug in the ET-VK `linear_dq8ca_q4gsw` kernel (negative per-token activation zero-point), producing garbled, runaway generation on Mali-G710 / G715. Weight-only int4 lowers to `et_vk.linear_q4gsw` (fp activation x int4 weight), bypassing the dq8ca path entirely and running correctly on Mali. This is a working Mali path / client workaround while the dq8ca kernel bug is fixed separately. Accuracy note: the TISO checkpoint was QAT-trained for 8da4w, so weight-only inference may be slightly less accurate than an accuracy-optimal weight-only-QAT model. ghstack-source-id: 398701430 @exported-using-ghexport Differential Revision: [D110111051](https://our.internmc.facebook.com/intern/diff/D110111051/)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Adds a
--weight-only-linearexport option that quantizes the TISO backbone's linears to weight-only int4 (dequantized int4 weights, fp activations) instead of the default 8da4w (dynamic int8 activations + int4 weights).Adds a
WeightOnlyInt4Linearmodule and_replace_linear_with_linear_int4_weight_only_for_pre_quantizationinexamples/models/llama/source_transformation/pre_quantization.py. It uses the same buffer layout asInt8DynActInt4WeightLinear(int8weightholding int4 values, per-groupscales/zeros), so an existing QAT checkpoint loads unchanged; its forward dequantizes the weight and runs a plainF.linearwith no per-token activation quantization.transform_linear_for_pre_quantizationgains aweight_onlyflag, threaded through the llama4 backbone build (backbone/model.py) and exposed as--weight-only-linearonexport_llm_backbone.py.Motivation: the 8da4w dynamic-activation-quant path triggers a Mali-GPU-specific numerical bug in the ET-VK
linear_dq8ca_q4gswkernel (negative per-token activation zero-point), producing garbled, runaway generation on Mali-G710 / G715. Weight-only int4 lowers toet_vk.linear_q4gsw(fp activation x int4 weight), bypassing the dq8ca path entirely and running correctly on Mali. This is a working Mali path / client workaround while the dq8ca kernel bug is fixed separately.Accuracy note: the TISO checkpoint was QAT-trained for 8da4w, so weight-only inference may be slightly less accurate than an accuracy-optimal weight-only-QAT model.
Differential Revision: D110111051